Distillation anomaly and fault detection based on clustering algorithms

■ Machine Learning for anomaly detection by modeling deviations from normal process. ■ Monitoring for predictive maintenance and fault detection based on Machine Learning. ■ Multiple clustering techniques to detect abnormal patterns in distillation units. ■ Validation in a chemical industry, with over 300,000 data collected for 3 years. ■ Proposed method applicable to other industrial processes and sectors. Anomaly detection in production processes is essential for ensuring reliability and efficiency in the industrial sector. In this way, system optimization requires advanced monitoring strategies such as predictive maintenance and intelligent fault detection. Traditional diagnostic methods rely on retrospective data analysis and deterministic cause-effect models, while machine learning approaches enable real-time monitoring and data-driven modeling to detect deviations from normal operation. This study proposes a scalable anomaly detection framework based on clustering algorithms, specifically applied to batch distillation processes—critical operations in chemical manufacturing that remain underexplored in real-world applications, particularly in multiproduct plants. The methodology was validated through an industrial case study at a chemical facility in El Palmar, Murcia (Spain), operated by a multinational corporation. Over 300,000 data points were collected over three years, focusing on critical variables governing distillation unit performance. Clustering techniques including k-means, DBSCAN, and hierarchical clustering were applied to identify deviations from standard operating conditions. Results demonstrate the effectiveness, flexibility, and scalability of the proposed approach, detecting anomalies in real time due to equipment faults, unstable conditions, or operator error. Integration of this system reduces unplanned shutdowns, improves energy efficiency, safety, and product quality, and provides operators with a real-time dashboard for decision support. Statistical evaluation of algorithms ensures adaptability across product types, while the custom application enables graphical monitoring of process deviations. Future work includes integrating performance indicators and ERP/MES connectivity. This framework serves as a reference model for deploying scalable anomaly detection systems across diverse industrial environments.